Inkling 2 Small: 594B MoE, open weights next week

Short answer

Thinking Machines Lab launched Inkling 2 Small on November 5, 2026: a multimodal Mixture-of-Experts model with 594 billion total parameters and 17 billion active, image input, and up to a 1 million-token context. You can try and fine-tune it now on Tinker. Full weights are planned under Apache 2.0 next week, after external safety testing.

What happened

On its research blog, Thinking Machines — founded by former OpenAI CTO Mira Murati — said Inkling 2 Small is the first model of its second generation. It takes text and images (no audio yet), supports variable reasoning "effort," and targets long-horizon coding, vision and knowledge work.

According to the company (and covered by MarkTechPost, Unite.AI and Artificial Analysis):

• Score: 31 on the Artificial Analysis Intelligence Index, up from 13 for Inkling Small.
• Where it runs today: Tinker Playground to try it; Tinker for inference and fine-tuning. Weights stay closed until the safety gate clears.
• What's coming: Apache 2.0 open weights next week if red-teaming finishes; other hosts later.
• Pricing on Tinker (Artificial Analysis): about $0.30 / $1.00 per million input / output tokens.

Demos lean agentic — interactive sites and games with visual feedback, plus an autonomous fine-tune of another small model. The lab also flags limits: it can be verbose, is weaker as pure chat, and sometimes overreacts on emotional-support prompts.

Why people are talking about it

Competitive US open-weight models are still scarce. Inkling 2 Small hit Techmeme hard, and on Hacker News it drew roughly 200 points and dozens of comments within hours — mostly about the MoE size, the Apache plan, and whether it is "good enough" for everyday knowledge work. For anyone who wants AI they can later customize or run themselves, this is a concrete next step.

What it means if you use AI

• Customization beats one-size chat. A base model you can fine-tune on Tinker (and soon download) lets you shape tone and skills for your own assistant workflow.
• Try before weights ship. Playground access today shows quality without waiting for the download.
• It is still a model. Strong demos do not make it a person; keep human review on anything that matters.
• Open weights change the privacy story. Once public, you or a host you trust control where it runs.

Quick checklist

1. Open Tinker Playground and run a short coding or vision task.
2. Note the effort setting: higher effort means slower, more thorough answers.
3. If you fine-tune, start small and keep a human review step.
4. Watch the Thinking Machines blog for the Apache 2.0 drop next week.
5. After weights land, choose Tinker, another host, or local — if your hardware allows.

Sources

• Thinking Machines Lab (Nov 5, 2026): Introducing Inkling 2 Small (primary)
• Thinking Machines Lab: Inkling 2 Small Model Card
• MarkTechPost (Nov 5, 2026): Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active
• Unite.AI (Nov 5, 2026): Thinking Machines Lab Launches Inkling 2 Small with Open Weights Pending
• Artificial Analysis: Inkling 2 Small index and pricing (via launch coverage)

Sources read on November 6, 2026 (Europe/Prague).

Related reading:

• Mistral Large 4 is now open: what it means for AI users
• DeepSeek Creator Mode turns a prompt into a plugin

Cover photo: Norbert Levajsics, "Laptop on a neat desk", via Wikimedia Commons (Unsplash), [CC0](https://creativecommons.org/publicdomain/zero/1.0/).